Thinking Machines’ new model takes the lead in western open-weight AI

Thinking Machines, a fledgling AI startup founded by former OpenAI Chief Technology Officer Mira Murati, has launched its newest model, Inkling 2 Small. According to independent testers, this model could potentially challenge the dominance of Chinese AI developers in the open-weight AI space. Until now, the model's weights were proprietary, but Thinking Machines announced they will release them next week.

Chinese companies have traditionally led the way in open-weight models, with DeepSeek and Moonshot frequently topping the leaderboards and matching the capabilities of pricier closed models from OpenAI and Anthropic. However, Inkling 2 Small, boasting 594 billion parameters and capable of processing images, appears to be a worthy contender. The model, which only generates text output, excels in web and game design, and showed significant improvements in coding abilities during testing.

Inkling 2 Small scored a 31 on the Artificial Analysis Intelligence Index, a composite test assessing progress towards artificial general intelligence. This score is above average for models in its class and shares the top spot with Gemini 3.8 Flash. For reference, top performers Gemini 4 Carbon and Claude Opus 5.5 scored 60 and 59, respectively.

The company's decision to release the model's weights next week is becoming increasingly common as AI companies become more security-conscious. This staggered process allows for more thorough testing by cybersecurity teams and safety reviewers before the weights are made public. Mistral, another French AI lab, recently released its Le Chonk model using the same approach, catching up to OpenAI's GPT-6.1 Terra in coding abilities.

Despite being priced at $0.30 for every one million input tokens, which is somewhat higher than the median cost of $0.24 for models in its class, Inkling 2 Small has shown impressive advancements. It has increased its automation capabilities by 41% and coding abilities by 20% compared to its predecessor, Inkling Small. Additionally, the model's accuracy has improved, with its hallucination rate decreasing from 61% to 27% and its accuracy rising from 32% to 34%.